New Automatic Allred Scoring for Breast Cancer Nuclei Detection and Scoring from ER-IHC Stained Images

Allred scoring is an important system for quantitatively evaluating estrogen receptor (ER) status and assisting pathologists in recommending hormone therapy. While several automatic methods have been proposed to overcome the high cost of manual counting, they face challenges such as expensive segmentation annotation costs, reduced classification accuracy due to loss of contextual information, and long evaluation times. This paper proposes an Integrated Segmentation and Classification model for Allred Score (ISCAS-Net) using an encoder-decoder network to provide more contextual information from patches for classification. A three-stage training approach is introduced to train the integrated model with only nuclei classification annotations, including unsupervised segmentation training, CNN classification model training with ground truth generation, and classification decoder training. A novel automatic Allred scoring system based on this model offers improved performance and significantly reduced evaluation time. Experiments on 37 immunohistochemically (IHC)-stained breast cancer Whole Slide Images (WSIs) show 83.78% agreement on hormone therapy recommendations and 59.46% agreement on Allred scores, both improving by 2.7% compared to the state-of-the-art (SoTA) method. Moreover, the processing time is reduced from 1.72 hours to 11 minutes per WSI. This system's faster scoring and improved performance have the potential to enhance clinical workflow and support more accurate hormone therapy recommendations.